Being cognitive is not running more agents. It is knowing which decisions are
settled, which are genuinely open, and which deserve a person.
The phrase usually gets attached to a control plane sitting over
a lot of agents: routing, queues, supervisors, retries, termination. That is real engineering and it
works. But it answers the question agentic orchestration already answered: who acts next.
Nothing in it is cognitive. It is a bigger diagram of the same thing.
What makes a person good at this is not that they decide
everything. It is that they know which mode they are in. Some things they run without thinking.
Some they hold open because the answer genuinely depends. And some they stop on, because they do not
recognize the situation. That is the most useful thing an expert ever says.
Almost every system fixes that allocation at design
time. An engineer draws the flowchart: this box is a rule, this box calls a model, this box needs
an approval. Then it is frozen. And it is wrong in both directions, every day, quietly.
So the argument here is narrow. Mode should be measured at runtime,
per decision, from what has already been decided. The instrument that measures it is the one
thing every pipeline currently discards as noise:
the times people disagreed.
Carey Chou ·
· rewritten Sep 29, 2026
01 · The pitch
The triage already happens. It just happens by accident.
Start somewhere unromantic. An ordering desk works an exception queue: a supplier
confirms short, a delivery slips, a cost comes back different from the purchase order. Somebody has
to decide: accept it, split the order, pay for faster freight, substitute, cancel.
1,200
exceptions in the queue on an ordinary day
18 h
attention available: three people, six usable hours each
30 h
attention needed, at ninety seconds to do one properly
−12 h
the gap. It does not close by working harder
So roughly
four in ten get no real thought. Nobody decided
which four.
✕
This is the part worth sitting with. The queue gets triaged either way: by
arrival order, by which ones look familiar, by who is tired at four o'clock. The skimmed ones are
not the easy ones. They are indistinguishable from the hard ones, because nothing in the
queue is labeled. So the choice was never whether to triage. It is whether to triage
on purpose.
What agentic orchestration offers here
Run the queue faster. Route it, retry it, escalate on a timeout. Genuinely useful,
and it does not touch the problem: more throughput on an unsorted queue just means skimming
more things sooner.
What is actually missing
A number, per item, saying how much thought this one deserves, plus the
honesty to say it about this case rather than about the category the case belongs to.
✔
To be clear about what comes first. Codifying the process (pulling the steps
and the written rules out of transcripts, procedures and documents) is the right first move, and
most teams have not finished it. This is about what you find after that is done: the
procedure is complete, the agent follows it correctly, and the queue is still being skimmed.
02 · What cognitive should mean
Three modes, and the whole job is knowing which one you are in
Not three kinds of task. Three states the same task can be in, depending on the
case in front of you. A short on a slow-moving item with a reliable supplier is one thing. The same
short, on the item in next week's ad, is a different decision wearing the same name.
Settledrun it
Comparable cases have come up before and they all went the same way.
There is no live question here. Acting costs nothing and asking costs a person ninety seconds
they need elsewhere.
Short of 3 units on an item with eleven weeks of cover. Accept, adjust the
order, move on. It has happened 44 times this quarter and it went the same way 41 of them.
Contestedhold it open
Comparable cases came up and people decided them differently. That is not
a gap in the data. It is the finding: this decision has a real trade in it, and it will never
become a rule without someone choosing what the trade should be.
Half the order short, on a promoted item. Pay for air freight, or substitute and
protect the margin? Past cases split 17 to 14, across all three buyers.
No precedentstop
Nothing close enough in the record to reason from. The right move is to say so.
A system that produces a confident answer here is not being helpful; it is
hiding the most important fact about the situation.
A supplier that has never missed, short on everything, the week after they
changed distributors. Three loose comparables, all more than a year old.
Agentic orchestration answers
who acts next. Codifying answers
what the steps are. This answers
how certain this one deserves to be.
Why a fixed allocation fails
Drawing the mode on a diagram is wrong in both directions
Over-determinized
The rule fires on a case it was never meant for. Nobody notices, because a rule
firing looks identical whether it was right or not. This failure scales exactly as well as the
automation does. That is the problem with it.
Over-escalated
A person is asked to approve something with one obvious answer, three hundred times
a week. They start approving without reading. The control still appears in the audit trail and
has stopped being a control. Worse, the attention it burned was the attention the hard
case needed.
◆
Both come from the same root. Somebody decided, once, at design time, what
kind of decision this is, and then the world kept moving while the diagram did not. The fix is not
a better diagram. It is to stop treating mode as a property of the box and start treating it
as a property of the case.
03 · The instrument
Reading ambiguity off the record
Here is the move the rest of this rests on. Every preference-learning pipeline built
in the last decade takes human decisions and fits a model to them. When two people decide the same
case differently, that is treated as label noise and averaged away. It is the standard,
sensible thing to do, and I think it throws away the most valuable signal in the data.
What disagreement is normally treated as
Noise. Inconsistency. Something to smooth out with more labels, tighter
guidelines, or a majority vote. The goal is a single reward model, so disagreement is an
obstacle to it.
What it actually is
The measurement. Qualified people, given the same case, deciding differently
is what ambiguity looks like from the outside. It is the only direct evidence you will ever get
that a decision contains a real trade rather than an unknown answer.
Agreement in the record says
this can be a rule. Disagreement says
this will never be one.
And it splits two ways, which matters a lot
Who disagreed tells you what to do about it
Between people
Buyer A expedites almost every time. Buyer B substitutes almost every time. Each of
them is internally consistent. This is not ambiguity; it is an unresolved policy question
that the organization has been re-litigating 260 times a day without noticing.
Escalate it once, get a ruling, write it down.
Within one person
The same buyer decides comparable cases differently, and cannot tell you why. That
almost always means your context is missing a variable they can see and you are not
recording. Go find it. It is usually something mundane: a note on the account, a
conversation, a date nobody joined to the record.
⚠
And the third case, which is the trap. Everyone agreeing does not mean they
are right. A whole desk can share a habit, and the record will show perfect consensus. So
settled means consistent, not correct. Those are two different claims, checked by two
different loops. Precedent tells you the mode. Only outcomes tell you whether the mode was
deserved.
04 · Confidence
A number that behaves the way intuition behaves
An expert's confidence is not a probability. It is recognition: have I seen
this before, and did it go the way I expect? That distinction sounds soft, and it has a very
practical consequence: recognition cannot be faked on a situation you have never seen.
A softmax can. A model will hand you 0.94 on a case with no precedent at all, and sound exactly
as sure as it does on its ten-thousandth easy one.
DensityHow many comparable cases exist in the record, and how close
they actually are.Few neighbors is low confidence, automatically. Nothing to fake it
with.
AgreementOf those, how many resolved the same way, and whether the
splits fall between people or inside one.This is the part that separates settled from
contested.
FreshnessWhen they were decided, and whether anything structural has
changed since.Forty agreeing cases from before the supplier changed hands are forty
cases about a different world.
# the readout, for one case in the queue
44 comparable cases · 90 days 41 of 44 agreed most recent 3 days ago → SETTLED
31 comparable cases · 90 days 17 / 14 split, all three buyers → CONTESTED
3 comparable cases · none inside 14 months → NO PRECEDENT
◆
The reason to build it this shape is that a person can argue with it.
"Forty-four similar cases, forty-one went the same way, the most recent was Tuesday" is a sentence
someone can push back on: those aren't similar, you're missing that two of them were
recalls. "Confidence: 0.87" gives them nothing to disagree with, which is not the same as
being right.
The comparison that decides everything
What counts as a comparable case
All three numbers depend on one choice: which past cases
count as near this one. Get it too loose and everything looks settled, because you are averaging over
situations that have nothing to do with each other. Too tight and nothing has precedent and the whole
queue escalates. This is the real engineering in the idea, and it is also the honest weak point;
see §10.
Start embarrassingly simple
Exact matches on a handful of fields people actually name out loud: item class,
supplier, size of the shortfall, whether the item is promoted, weeks of cover. A lookup, not
a model. It is inspectable, and it is usually enough to sort most of a queue.
Earn the complexity later
Learned similarity is better and it costs you the ability to explain a verdict. Buy
that trade only where the simple version measurably fails, and keep the simple one running
beside it, because you will need something to show the person you just interrupted.
05 · Before any framework
One exception, turn by turn
Assume the easy part is done: the process is codified, the rules are written, the
agent applies them correctly. A contested case arrives. What changes is not the answer; it is
that the system says which mode it is in first, and that turns out to change the whole
exchange.
AI
Supplier confirmed 48 of 100 cases on the promoted item. I'm not going to pick for
you. Comparable cases split 17 to 14, and the split runs across all three of you, so
this one has a real trade in it. Here are both sides.
names its mode
AI
Expedite the balance: covers the ad, costs about $2,900 in freight.
Substitute the pack size: no freight, roughly 14% of the forecast demand walks.
shows the trade
YOU
Expedite. We can't be short on an ad item.
decision
◆
Nothing has gone wrong yet, and one thing has already gone right. The system
did not recommend and then get overruled. It said this is contested and handed over the
trade instead of the answer. That costs the person the same ninety seconds, but they spend it
deciding rather than auditing.
The next two turns are where the
durable part comes from.
AI
Would that still hold if it weren't an ad item?
one variable
YOU
No. Off-ad I'd substitute every time and never think about it.
the boundary
AI
And if the freight were $8,000 instead of $2,900?
one variable
YOU
Still expedite. Being short on an ad is worse than the freight number, almost
whatever the freight number is.
the weight
The technique in those four turns
Change one thing and ask whether the answer survives
Notice it never asked why. Asked why, people build a
reason after the fact, fluently, in the language of the official objective, and it is not what drove
the choice. Asked whether a decision survives one changed variable, people answer accurately and
instantly, because where a rule stops is something they know precisely even when they cannot state
the rule.
"Same short, but the item isn't on ad?"
flips
The weight sits on promotional exposure, not on the shortfall size.
"Same case, but freight is nearly three times as much?"
holds
The ad constraint dominates freight cost across a wide range. That is a weight, and
it was never written down.
"Same case, but eleven weeks of cover on hand?"
flips
It is coverage doing the work, not the promotion by itself. Two variables, not one.
"Would you have said this before the stockout in March?"
flips
Part of this weight is a recent bad month, not a standing preference. Flag it and
re-ask in a quarter.
⚑
The last probe is the one everyone skips. A counterfactual across
time is the only cheap way to separate a lasting preference from a reaction to something
recent. If changing only the date flips the answer, you were about to write down a mood as a
rule, and then apply it for years after the person themselves moved on from it.
✔
The cost matters as much as the technique. Four questions, four sentences,
inside a decision the person was making anyway. No workshop, no interview, no form. If getting a
weight out of somebody costs more than that, it will get done once during the pilot and never
again.
06 · What you already have
Mining, bots, and people: three parts of one loop
None of this needs a new platform, and that is most of the argument for it. The three
capabilities involved are usually already bought, already running, and pointed at three unrelated
goals by three different teams.
Process mining is the memory
It reads the event log and tells you what actually happened, as opposed to what the
documented procedure says. Every deviation it finds is an override somebody already made,
timestamped, with a case attached. That is the precedent store: density, agreement and freshness
all compute directly off it. You are not collecting new data. You are reading data you keep for
compliance and asking a different question of it.
RPA is the hands, and the honest test
The bots execute the settled cases. More usefully, they falsify your mode
assignment: if a case is truly settled you can hand it to a bot, and if you keep needing a
person "just to glance at it" then it was never settled and your routing is lying to you.
Automation also buys volume, and volume is what makes drift visible as a statistic rather than
as an anecdote.
Human attention is the budget
Not a rubber stamp at the end of a flow. The scarce input the whole system exists to
allocate. Which reframes the success metric: not accuracy, but decisions per hour of expert
attention at equal or better outcomes. That is a number a business will actually fund, and one you
can put against the 18 hours from §01.
Mining says
what we have done. Bots say
what we can stop doing. People decide
what is left.
The arithmetic, once it is sorted
The same queue, the same three people
852
settled · bots run them · zero attention
264
contested · trade shown, person decides
84
no precedent · stopped, logged carefully
8.7 h
attention needed, against 18 available
◆
Those proportions are illustrative; the point is the shape, not the numbers.
The deficit in §01 was never a staffing problem. It was 1,200 items arriving with no indication of
which ones were questions. Sorting the queue bought more than hiring would have, and it
bought it out of logs the company was already keeping.
07 · The part nobody budgets for
Automate a decision and it stops teaching you anything
This is the consequence that makes the whole thing an engineering problem rather than
an idea, and it is easy to miss because it looks like success.
✕
A case gets classified settled. The bot takes it. From that moment,
no human ever decides one of those again, so no new precedent is created, agreement stays
frozen at whatever it was on the day you automated, and confidence stays high forever. The world
moves. The measurement cannot move with it, because you removed the only thing that was measuring.
The system does not get it wrong. It goes blind, while reporting that it is certain.
What it looks like from the dashboard
Automation rate up. Escalations down. Confidence high and stable. Every number you
would have picked as a KPI moving in the direction you wanted, for months after the
underlying decision stopped being correct.
The only fix, and it is unglamorous
Deliberately route a small sample of settled cases back to people anyway:
cases you are confident about, that the bot could have handled. You are buying precedent you
would otherwise stop generating. It is a holdout, and it is the price of the confidence number
staying true.
# the sampling budget, on the numbers from §06
sample 3% of settled 26 cases / day
at 90 seconds each 39 minutes / day
against attention available 18 hours
# what it buys
drift on an automated decision becomes visible in weeks, not at the next audit
⚑
It will look like waste, and you should expect to defend it. You are paying
people to decide things a bot already knows how to decide. That is the same argument as a holdout
group in an experiment, and it loses the same way every time it loses:
someone reclaims the budget because nothing has gone wrong yet.
And it runs the other way too
The boundary is supposed to move
A contested decision does not stay contested. Once the policy question underneath it
gets ruled on (once somebody with the authority says on ad items, we eat the freight), the
splits stop, agreement climbs, and the case becomes settled on its own. The line between deterministic
and ambiguous is not a design artifact. It is a measurement, and it should look different this
quarter than it did last one. If it never moves, nothing is being learned, and the routing has
quietly become another frozen diagram.
08 · Where the contested ones go
What a contested case leaves behind
Only contested cases need any of this. That is worth saying plainly, because these
projects usually begin by trying to build one reward model over everything. You do not need a
reward model for a settled decision; precedent is already enough. Spend the modeling
effort on the fifth of the queue where there is an actual question.
Lands in the system of record
Expedited freight approved, $2,900
Date, order, approver
Both true. Neither tells you anything about how to handle the next one.
Said once, then gone
Freight cost was not the deciding factor
Promotional exposure was, and it is not in the procedure
It only applies while the item is on ad
It got this heavy partly because of one stockout in March
Four things, none recorded anywhere, all of them needed to get the next one right.
Implicit weightwhich goal gave way, and by how muchfreight cost → not the deciding factor
Boundarywhere the preference stops applyingon-ad items only, and only under four weeks of cover
Tacit constrainta limit never expressed, because everyone knew itnever short on an ad item, whatever it costs
Sub-goala stand-in being optimized instead of the real goal"never short on ad" standing in for "the promotion pays back"
And then the uncomfortable part
What you learn is a hypothesis, not a rule
Codifying assumes what the expert tells you is correct and
worth writing down. That is a fair assumption for steps and a poor one for weights. Some of what you
pick up will be wrong, and a system that only absorbs will scale that up faithfully. So every learned
criterion gets checked against the goal it claims to serve, and there are three outcomes:
Promoteevidence links it to the goal
It measurably helps the objective above it. It becomes a written criterion with
a weight, a scope and a version. And a date.
Challengeit works against the goal
Real, consistently applied, and costing more than it protects. Price it in
the goal's own units and put the number in front of someone who can overrule it. With a number
attached it is a decision; without one it is an argument about preferences.
Retireright answer, old question
It was correct when it formed and the conditions changed. Nobody revisits these,
because they never felt like decisions; they feel like how things are done.
◆
Same criterion, all three verdicts, depending only on the evidence. "Never short on
an ad item" is a promote if the promotional lift data backs it, a challenge if the
freight has been costing more than the stockouts ever did, and a retire if it came from one
bad quarter three years ago that nobody has looked at since.
09 · Approach & techniques
Built from parts that already exist
Almost none of this needs new science. The work is wiring mature things together and
pointing them at a question they are not normally asked, plus one step the standard toolkit
genuinely does not cover.
1
Precedent store
conformance checking · deviation mining
2
Similarity
exact-match strata, then learned
3
Mode routing
density · agreement · freshness
4
Expose the trade
Pareto front, not a scalar
5
Infer weights
preference models · inverse RL
6
Sampling budget
the missing one
Step
What already exists
The part you have to add
Precedent store
Process mining, conformance checking. Mature, and usually already licensed.
Nothing technical. You are asking it "where did people deviate" rather than
"where are we non-compliant." Same query, opposite intent.
Similarity
Stratification, nearest neighbors, learned metrics. All standard.
The fields. Which variables define "a comparable case" is a domain question, and it is
the single biggest determinant of whether any of this works.
Mode routing
Conformal prediction gives a set with a coverage guarantee and no distributional
assumptions. Learning-to-defer turns that into a policy. Both mature.
Neither is normally wired to disagreement among past humans, or to how reversible
the action is. Those two are what make the routing behave sensibly.
Expose the trade
Multi-objective optimization. Fifty years old.
Resisting the urge to collapse it. A weighted score picks one point on the frontier and
presents it as the answer; somebody chose those weights and nobody wrote it down.
Infer weights
Bradley–Terry, DPO, inverse RL, conjoint analysis, Bayesian elicitation. Deep literature.
Run it only on contested strata, and stop averaging over disagreement: fit per
decision-maker where the split is between people.
Sampling budget
Bandits and holdout design have the math. It is not new.
Nobody applies it to automated decisions. The whole literature is about exploring
actions, not about re-sampling decisions you already trust to keep them measurable.
⑥
Step 6 is the one with no off-the-shelf answer. Preference learning is built
to fit human choices, and fitting them better is how it gets scored. Nothing in the standard
pipeline asks whether the preference it just learned is still being observed, or whether it should
be followed at all.
# what the routing actually computes, per case
near = { past cases within the similarity strata of s }
conf = f( |near| , agreement(near) , recency(near) )
# and the two things that make it behave like judgment
agreement splits between people → unresolved policy · escalate once
agreement splits within a person → missing variable · go find it
10 · Where this could be wrong
The claims most likely to break
Similarity is load-bearing
Everything rests on "comparable case", and that definition is a judgment call
smuggled in at the bottom of a framework about surfacing judgment calls. Draw it loosely and the
whole queue reads settled.
Consensus can be habit
Agreement measures consistency, not correctness. A desk that has been wrong the same
way for four years produces a beautiful settled signal, and automating it scales the error.
Thin records
Plenty of decisions are made in a conversation and never written anywhere. No log,
no precedent, no routing. And the decisions that work that way are often the important ones.
The holdout gets cut
It is the first budget line to go, it will be cut during a good quarter, and the
cost of cutting it does not show up until well after whoever cut it has moved on.
⚖
The test, before building anything. Take a month of decisions already made.
Compute density, agreement and freshness for each, and route them on paper. Then check the
ones the routing called settled: did they actually go the way it predicted? If a routed
queue does not beat arrival order at picking which decisions deserved attention, the sorting was
never the bottleneck, and this is an expensive way to draw a nicer flowchart.
What this makes it
A thought partner is not a system that is always talking
The phrase gets used loosely. Mode routing gives it
something precise to mean, and it is mostly about restraint:
On settled cases
It acts, and does not bother you. A partner who asks about everything is not a
partner, they are a queue with manners.
On contested cases
It frames the trade instead of handing over an answer to audit. You spend the same
ninety seconds deciding rather than checking.
On cases with no precedent
It says "I don't recognize this", the single most useful thing a colleague
ever says, and the thing these systems almost never do.
◆
Which is the honest answer to what makes it cognitive. Not more autonomy, and
not more agents under one control plane. Knowing which of the three you are in, measuring it
from what has already been decided, and letting that measurement move.
